Even in 2019, there is still a lot of handwriting done by doctors, simply because not all use cases can be digitized easily. For the Flemish Government, this means that humans still need to read doctor's handwritings and enter the text in databases through custom-made data-entry form applications. This is a time-consuming process, especially for medical data, which needs to be treated with great accuracy. Also, a fraction of about 10% of handwritings cannot be read with sufficient accuracy by a human, and requires double-checking.
We developed a data-processing pipeline for one such use case with the goal of supporting the humans by reading the handwritings and predicting what is written. Together with the predicted text, we calculate confidences for the predictions and we show that we can define confidence levels above which the machine predictions perform better than a human. We also show that the machine can suggest predictions for illegible handwritings, which allows the human to decreases the number of illegible handwritings by about 30%.
In this presentation, we will present the end-to-end pipeline, from data mining and anonymisation, through data cleaning with numerous image-processing steps, to building and applying the deep-learning model, augmented with some natural language processing.
We will show real-life examples and present ample statics to show the validity of the model. This type of AI modelling of doctor's handwriting can be applied to numerous related use cases.
Educated with a PhD in astronomy, moved into IT around 1998, and since then always done innovative projects with the latest technology. Moved from Java and SOA integration architecture into Big Data architecture and data science, now into AI.
|Talks by tracks||Talks by session types||List of Speakers||Schedule|